大型语言模型衍生的重新情境化揭示跨癌症的功能图景

The vast heterogeneity of human cancer necessitates a deeper understanding of how fundamental biological pathways are functionally reconfigured across different cancer types in contrast to non-malignant diseases. Conventional pathway analyses based on static, disease-agnostic gene sets often obscure the plasticity and tumor-specific roles of pathways. We address this limitation by leveraging the knowledge integration of large language models (LLMs) to generate and embed context-aware pathway descriptions, enabling a novel framework for quantifying pathway relevance and functional variability across the oncological landscape. We applied this framework to analyze 268 KEGG pathways across 35 cancer types and 7 non-malignant diseases. We performed two complementary analyses of the LLM-generated context-aware pathway descriptions: 1) disease-level analysis by applying unsupervised clustering to the embeddings derived from the context-aware descriptions to map inter-disease functional relationships; and 2) pathway-level analysis that quantifies the functional variability of each pathway across diseases to understand their disease-specific roles. The disease-level analysis revealed a coherent functional atlas of diseases. Within cancers, lineage and cell type drove distinct yet biologically consistent groups, such as hematologic malignancies, gastrointestinal cancers, and hormonal cancers. Non-malignant conditions (e.g., neurodegenerative disorders) formed separate, distinct clusters. Pathway-level analysis showed marked functional variability. Pathways with the lowest variation across diseases mainly involved conserved cellular functions (e.g., mitophagy, AMPK signaling ), while highly dispersed pathways captured context-specific programs (e.g., viral carcinogenesis, transcriptional misregulation in cancer ). Independent validation using a literature-derived pathway relevance score across diseases confirmed that the top 20 most dispersed pathways had significantly higher relevance scores than the bottom 20 (p < 0.01). Furthermore, the continuous level of dispersion correlated positively with pathway relevance (p < 0.01). These observations support that the semantic heterogeneity captured by LLMs across diseases truly reflects biological specificity and functional relevance. In summary, LLM-generated context-aware pathway descriptions and their corresponding embeddings successfully capture disease-specific functional organization and reveal mechanistic coherence across cancers. This represents a significant methodological advance over traditional static approaches, providing a dynamic, biologically relevant map for understanding heterogeneous cancer mechanisms and identifying novel therapeutic strategies.